ColossalAI/colossalai/utils/activation_checkpoint.py

150 lines
5.5 KiB
Python
Raw Normal View History

2021-10-28 16:21:23 +00:00
#!/usr/bin/env python
# -*- encoding: utf-8 -*-
import torch
from torch.utils.checkpoint import check_backward_validity, detach_variable
from colossalai.context.random import get_states, get_current_mode, set_seed_states, set_mode, sync_states
from .cuda import get_current_device
2021-10-28 16:21:23 +00:00
def copy_to_device(obj, device):
if torch.is_tensor(obj):
# Notice:
# When in no_grad context, requires_gard is False after movement
ret = obj.to(device).detach()
ret.requires_grad = obj.requires_grad
return ret
elif isinstance(obj, list):
return [copy_to_device(i, device) for i in obj]
elif isinstance(obj, tuple):
return tuple([copy_to_device(v, device) for v in obj])
elif isinstance(obj, dict):
return {k: copy_to_device(v, device) for k, v in obj.items()}
else:
return obj
2021-10-28 16:21:23 +00:00
class CheckpointFunction(torch.autograd.Function):
@staticmethod
def forward(ctx, run_function, activation_offload=False, *args):
2021-10-28 16:21:23 +00:00
check_backward_validity(args)
ctx.run_function = run_function
ctx.activation_offload = activation_offload
ctx.device = get_current_device()
2021-10-28 16:21:23 +00:00
# preserve rng states
ctx.fwd_cpu_rng_state = torch.get_rng_state()
sync_states()
ctx.fwd_seed_states = get_states(copy=True)
ctx.fwd_current_mode = get_current_mode()
if hasattr(torch, 'is_autocast_enabled'):
ctx.had_autocast_in_fwd = torch.is_autocast_enabled()
else:
ctx.had_autocast_in_fwd = False
if activation_offload:
inputs_cuda = copy_to_device(args, ctx.device)
else:
inputs_cuda = args
2021-10-28 16:21:23 +00:00
with torch.no_grad():
outputs = run_function(*inputs_cuda)
2021-10-28 16:21:23 +00:00
# Save non-tensor inputs in ctx, keep a placeholder None for tensors
# to be filled out during the backward.
ctx.inputs = []
ctx.tensor_indices = []
tensor_inputs = []
for i, arg in enumerate(args):
if torch.is_tensor(arg):
if activation_offload:
tensor_inputs.append(copy_to_device(arg, 'cpu'))
else:
tensor_inputs.append(arg)
2021-10-28 16:21:23 +00:00
ctx.tensor_indices.append(i)
ctx.inputs.append(None)
else:
ctx.inputs.append(arg)
if activation_offload:
ctx.tensor_inputs = tensor_inputs
else:
ctx.save_for_backward(*tensor_inputs)
2021-10-28 16:21:23 +00:00
return outputs
@staticmethod
def backward(ctx, *args):
if not torch.autograd._is_checkpoint_valid():
raise RuntimeError("Checkpointing is not compatible with .grad() or when an `inputs` parameter is "
"passed to .backward(). Please use .backward() and do not pass its `inputs` argument.")
2021-10-28 16:21:23 +00:00
# Copy the list to avoid modifying original list.
inputs = list(ctx.inputs)
tensor_indices = ctx.tensor_indices
if ctx.activation_offload:
tensors = ctx.tensor_inputs
else:
tensors = ctx.saved_tensors
2021-10-28 16:21:23 +00:00
# store the current states
bwd_cpu_rng_state = torch.get_rng_state()
sync_states()
bwd_seed_states = get_states(copy=True)
bwd_current_mode = get_current_mode()
# set the states to what it used to be
torch.set_rng_state(ctx.fwd_cpu_rng_state)
for parallel_mode, state in ctx.fwd_seed_states.items():
set_seed_states(parallel_mode, state)
set_mode(ctx.fwd_current_mode)
if ctx.activation_offload:
tensors = copy_to_device(tensors, ctx.device)
2021-10-28 16:21:23 +00:00
# Fill in inputs with appropriate saved tensors.
for i, idx in enumerate(tensor_indices):
inputs[idx] = tensors[i]
2021-10-28 16:21:23 +00:00
detached_inputs = detach_variable(tuple(inputs))
if ctx.had_autocast_in_fwd:
with torch.enable_grad(), torch.cuda.amp.autocast():
outputs = ctx.run_function(*detached_inputs)
else:
with torch.enable_grad():
outputs = ctx.run_function(*detached_inputs)
if isinstance(outputs, torch.Tensor):
outputs = (outputs,)
# recover the rng states
torch.set_rng_state(bwd_cpu_rng_state)
for parallel_mode, state in bwd_seed_states.items():
set_seed_states(parallel_mode, state)
set_mode(bwd_current_mode)
# run backward() with only tensor that requires grad
outputs_with_grad = []
args_with_grad = []
for i in range(len(outputs)):
if torch.is_tensor(outputs[i]) and outputs[i].requires_grad:
outputs_with_grad.append(outputs[i])
args_with_grad.append(args[i])
if len(outputs_with_grad) == 0:
raise RuntimeError("none of output has requires_grad=True,"
" this checkpoint() is not necessary")
2021-10-28 16:21:23 +00:00
torch.autograd.backward(outputs_with_grad, args_with_grad)
grads = tuple(inp.grad if isinstance(inp, torch.Tensor) else None for inp in detached_inputs)
return (None, None) + grads
2021-10-28 16:21:23 +00:00
def checkpoint(function, activation_offload, *args):
2022-03-25 05:02:39 +00:00
"""Checkpoint the computation while preserve the rng states, modified from Pytorch torch.utils.checkpoint.
2021-10-28 16:21:23 +00:00
2022-03-25 05:02:39 +00:00
Args:
function: Describe the forward pass function. It should know how to handle the input tuples.
args (list): Tuple containing the parameters of the function
Returns:
Output of running function with provided args.
2022-01-21 02:44:30 +00:00
"""
return CheckpointFunction.apply(function, activation_offload, *args)